Fig. 4

Dissecting the performance of MIOSTONE through control studies. a MIOSTONE demonstrates robustness across various taxonomic trees. Two variations of MIOSTONE, utilizing taxonomies from GTDB and NCBI respectively, demonstrate comparable predictive performance across seven real microbiome datasets. b While MIOSTONE can emulate any hierarchical correlation among taxa within its architecture, alternatives, such as phylogenic trees, perform significantly worse than the taxonomy-encoding architectures. c MIOSTONE’s data-driven aggregation of neuron representations either outperforms or matches the performance of the deterministic selection of nonlinear representations across most datasets. d By assigning larger taxonomic groups greater representation dimensionality, MIOSTONE can capture more complex biological patterns for trait prediction, outperforming methods that use fixed representation dimensionality. e MIOSTONE demonstrates robustness in selecting the hyperparameter that controls taxonomy-dependent representation dimensionality. f The curse of dimensionality cannot simply be mitigated using feature selection. MIOSTONE trained with all microbiome features, either outperforms or matches the performance of the model trained with a subset of highly variable taxa across most datasets. All settings used by MIOSTONE are marked by \(\bigstar\)